Parsing Molecules for Drug Discovery.

Current events are painfully reminding us of our vulnerability to infectious disease. Most readers will not need to be reminded that the effectiveness of our currently used antibacterial agents is diminishing due to growing bacterial resistance, and developing new agents by modifying currently used agents has failed to fully address our needs. As a result, the search for new agents, especially ones with new chemotypes and targets, has become increasingly urgent.

Recently, Stokes et al. reported a deep learning approach to discover antibacterials that would meet these requirements.[1] Their approach, which is conceptually simple and technically challenging, required three things: (1) a training set, molecules with known structures and known ability to either stop or not stop bacterial growth, (2) a program to “learn” from this data set [in this case, the program was Chemprop, a directed message passing neural network (more on this below)], and (3) a screening set, a large number of structures from which Chemprop could identify likely antibacterials.

The authors began by constructing the training set. Rather than relying on previously constructed sets or trying to assemble a data set from the literature, the authors created their data set by experimentally screening 2335 compounds with known structures for their ability to inhibit the growth of Escherichia coli. The molecules in this initial training library consisted mostly of Food and Drug Administration-approved drugs but also included natural products isolated from plant, animal, and microbial sources. Because the screen is blind to the mechanism of action and there are many ways to inhibit bacterial growth, positive screening hits will work through different mechanisms. In principle, a model trained on this data set could predict activity for molecules that act through different and possibly new mechanisms of action.

The authors then used the training data set from their phenotypic screen to train the directed message passing neural network (DMPNN), Chemprop, to predict which compounds are likely to inhibit the growth of E. coli. The term parsing may remind readers of long ago grammar classes in which sentences were parsed (verb, noun, adjectives, etc.). Chemprop parsed molecules into smaller fragments [amide bond, aromatic ring, methyl group, etc. (Figure )]. A directed message passing neural network can learn which substructures of a molecule are associated with a given property, in this case the ability to inhibit the growth of E. coli. DMPNNs can learn substructures of molecules by passing “messages”, which contain information about the bonds and atoms, around local regions of the molecule. Through message passing, the neural network could identify functional groups and other local features. Because DMPNNs distinguish only local features of the molecule, Chemprop also uses features of the entire molecule computed using RDkit, an existing toolkit for computing molecular properties, such as the number of hydrogen bond donors and acceptors the and number of rotatable bonds, to make its predictions. Previously, Chemprop had outperformed many other machine learning methods for predicting chemical properties.[2]

Figure 1

Chemprop correctly predicted that halicin could inhibit the growth of E. coli after it was trained on the results of a phenotypic screen. Chemprop makes its predictions using a directed message passing neural network that learns how local features are associated with the ability of a molecule to inhibit the growth of E. coli.

Deep learning approaches have the potential to increase the efficiency of high-throughput screening. While high-throughput functional screening can screen hundreds of thousands to a million compounds for a given activity, those numbers comprise only a tiny fraction of small molecule space. A deep learning computational screen can make predictions for at least 108 compounds, many multiples of what could be experimentally screened, and a computational screen is not limited to available molecules. However, models based on deep learning are not sufficiently accurate to pick out the most promising compound from 108 compounds, but they can prioritize compounds for functional screening with much higher hit rates. This ability to focus screens on the most likely molecules has obvious advantages for drug discovery.

Stokes et al. demonstrated Chemprop’s ability to increase the efficiency of screens for antibiotics by applying it to three different compound libraries, the Drug Repurposing Hub, the WuXi antituberculosis (https://www.broadinstitute.org/therapeutics-platform/compound-libraries), and the ZINC15 libraries (http://zinc15.docking.org/).[3] The model was exceptionally accurate for the Drug Repurposing Hub, with a true positive rate of 51.5% for the top 99 ranked molecules and a false negative rate of only 3.2% for the 63 lowest ranked molecules. The much higher hit rate for the top-ranked molecules illustrates deep learning’s ability to prioritize molecules for screening. From the set of the top-ranked molecules, the authors zeroed in on a synthetic small molecule that was previously identified as a selective c-Jun N-terminal kinase inhibitor. The small molecule (SU3327, renamed halicin) showed a micromolar minimal inhibitory concentration (MIC) against a range of bacteria, including high-priority pathogens such as Acinetobacter baumannii, Staphylococcus aureus, Clostridium difficile, and carbapenem-resistant Enterobacteriaceae. Halicin transiently reduces the level of proton polarization of the bacterial membrane. While a proton gradient is restored within several minutes, cell death does not occur until hours later; therefore, the exact mechanism of action remains to be determined. Nevertheless, halicin’s killing of dormant bacteria distinguishes it from most antibiotics, which kill actively growing bacteria and could be used to eradicate recalcitrant persisters. Importantly, the molecule displays promising activity in two different animal infection models against divergent pathogens, further suggesting its potential as a therapeutic agent. Side effects remain to be established.

After their success with the Drug Repurposing Hub library, the authors applied their deep learning approach to two other compound libraries, the WuXi antituberculosis and the ZINC15 libraries, both of which cover wider swathes of chemical space than either the training library or the Drug Repurposing Hub library. Their model predicted that it was unlikely that any molecules in the WuXi library were active and none of the top 200 or bottom 100 ranked molecules in the WuXi library were active against E. coli when assayed. A small group of 23 high-scoring molecules from the ZINC15 were assayed for activity against E. coli, and eight exhibited some degree of activity, a true positive rate of 35%, considerably lower than the Drug Repurposing Hub true positive rate of 51.5%. One possible explanation for this difference would be the Drug Repurposing Hub’s occupying a region of chemical space more similar to the training set than the ZINC15 library. Of the eight active compounds from the ZINC15 library, only two had MICs of <10 μg/mL against E. coli. However, it is not surprising that their model is not very predictive of MICs as it was trained to predict activity at a single relatively high concentration (50 μM). Both of these observations highlight the crucial role the training set plays in making accurate predictions and suggests that with a more diverse training set Chemprop would be able to predict antibiotic activity even more accurately. While further tests of Chemprop on more diverse libraries are required to determine the true extent to which deep learning can improve the efficiency of high-throughput screening, this initial report hints at the great promise deep learning holds for the discovery of novel antibiotics.